Workflow
数据炼金术(DataAlchemy)
icon
Search documents
链式思维是幻象吗?从数据分布视角重新审视大模型推理,马斯克回复,Grok破防
机器之心· 2025-08-14 09:11
Core Viewpoint - The research suggests that Chain-of-Thought (CoT) reasoning in large language models (LLMs) may not represent true reasoning but rather a replication of patterns learned from training data, leading to fragility when faced with out-of-distribution tasks [2][10][37]. Data Distribution Perspective on CoT - The effectiveness of CoT is attributed to the "structured inductive bias" learned within the training distribution, indicating that the reasoning chains are merely reproductions of common patterns rather than genuine logical deductions [13][37]. - A theoretical framework is introduced to quantify the relationship between training and testing distributions, highlighting how distribution shifts can impact reasoning performance [15]. Experimental Findings on Generalization - In "task generalization," the model shows nearly 100% accuracy within the training distribution, but accuracy drops to 0.01% with slight distribution shifts, indicating a lack of true generalization [23]. - Supervised fine-tuning on a small amount of new data can restore performance, but this only expands the existing distribution boundaries without enhancing abstract generalization capabilities [24]. - In "length generalization," even minor changes in input sequence length significantly affect model performance, demonstrating a tendency to generate reasoning chains consistent with training lengths [26]. - The model is highly sensitive to format changes, with even minor alterations in input prompts leading to complete reasoning failures [28]. Universal Sensitivity to Distribution Shifts - The study finds that the sensitivity to distribution shifts is a common phenomenon across different sampling temperatures and model sizes, indicating that this issue is not isolated to specific models [31]. Practical Implications - In high-risk fields such as healthcare and finance, reliance on CoT for robust reasoning is cautioned against, as misleading reasoning chains can be more dangerous than outright incorrect answers [34]. - Current evaluation methods that depend on validation sets closely aligned with training distributions may overestimate model robustness, necessitating stricter out-of-distribution testing [35]. - While supervised fine-tuning can quickly enhance performance on specific tasks, it does not equip models with true abstract reasoning capabilities [36].